problem about the reproduce of MOTCVPR19 results

[yjmlaile]

hi thanks for your sharing of code. after conducting experiment with unchanged configuration on the MOTCVPR19 dataset, I get a bad results for train part: MOTA MOTP FN FN ID_SW
3.895% 92.627% 1082296.0 7889.0 235.0 I want to know if the given hyper-parameters are different from that used in your experiment.

[fangxiaoran]

Hi, I'm curious about how you install the KCF2 module, cause iou-tracker needs a python3 env while KCF2 needs python2 as describe. How do you solve it? Thx. btw, I got a bad result like yours if I use KCF rather than KCF2.

[bochinski]

@yjmlaile do you get by any chance a message like "[warning] KCF2 not available, falling back to KCF. please see README.md for further details" ? that means you should install the alternative (better) implementation of the KCF tracker. The version from opencv does not perform well (at least from my experience) but it is easier to use.

@fangxiaoran you can safely ignore the python2 requirement, it should work anyway

[fangxiaoran]

@bochinski Thank you it works. 😊

@yjmlaile Do you know where I can download the CVPR 2019 dataset now? I can't find it on the official website.

[yjmlaile]

@fangxiaoran The CVPR 2019 dataset is denoted as MOT20 which can be downloaded in MOT20

[fangxiaoran]

@yjmlaile Thanks! But I think there is a little difference in sequences' resolution. Anyway, do you have a latest result? Mine still bad...

[yjmlaile]

@fangxiaoran I can't install the KCF2, so my results are still bad. Do you install the KCF2 successfully?

[fangxiaoran]

@yjmlaile Yes, I can install KCF2. What's your problem?

[yjmlaile]

@fangxiaoran I run the setup.py of its source code, with meeting its requirements, in a windows system. I get a error 'LINK : fatal error LNK1181: cannot open the input file“opencv_core.lib”'. Then I install the KCF2 in a linux system and get a error 'cannot find -lopencv_core collect2: error: ld returned 1 exit status error: command 'g++' failed with exit status 1'. I have no idea about that and I want know your requirements configuration and installation process, especially the interface of opencv.

[fangxiaoran]

@yjmlaile I tried on Linux and built OpenCV manually as the official tutorial. The version is 3.4.10. I also installed opencv-python module by pip. The version is 3.4.9.31 because I can't find a 3.4.10 version. Can you find libopencv_core.so in your machine? If so, maybe you can try this.

[yjmlaile]

@fangxiaoran @bochinski Thanks for your advises, after installing the KCF2 and running the experiment, I get a marginal improvement of results on MOT20 training dataset: MOTA MOTP FP FN IDsw 4.101% 92.153% 8254 1079596 235 So maybe the given hyper-parameters are different from the those using on public results. I think even there is difference between MOT20 and MOTCVPR2019 dataset, the results is so bad when comparing with the public results.

[fangxiaoran]

@yjmlaile Glad to see you get it done. My results is almost the same.